Dated observation note. Recorded 2026-08-10 against the frozen dataset datasets/DS-004-entity-clustering-v1/ (generated 2026-08-09T22:45:56Z at snapshot tip 97,734, hash ebb3107d2ad9edfd6e61217eaaf5c6e7951d874ee75a64a710d30dab664f581e, REPEATABLE READ). Method: keshictl cospend-cluster (cmd/keshictl/cospend.go, tool version 41310cac02bf-wb1bbc97). The run and the finding were both independently reviewed; the fleet decomposition in §3 was independently reproduced by the finding review. Committed context: Keshi's internal roadmap §14.1 (FULL RUN DONE 2026-08-10).

Status: internal research record, exploratory (not pre-registered). Publication is a separately held decision; nothing here is written for release. §2 separates what is frozen in DS-004 from what is not.

Status revision, 2026-08-11. The held-publication sentence above is superseded: an operator decision re-scoped publication to the site package, and this note ships inside it. The exploratory, not-pre-registered status and every boundary in §5 are unchanged.

Unit of analysis throughout: blocks with a decoded coinbase payout address. The snapshot holds 97,735 canonical blocks up to the bound, of which 97,734 carry a decoded coinbase payout address; only those enter every share and coefficient below, on both bases.

Summary sentence, with the era qualifier it must always carry: on the all-time cumulative basis, the second-largest block-producing entity in the chain's history is an unattributed genesis-era fleet of 107 uniform workers that mined roughly 29% of blocks up to height 30,768, went dark, and still holds about 27.6M PRL; address-basis figures hide it, and the full-network Nakamoto coefficient falls from 50 (addresses) to 12 (entities) on its account alone.

1. The finding

Method in one line: union-find every set of addresses that co-spend in a non-coinbase transaction into one entity (the standard multi-input heuristic; coinbase inputs excluded), over one repeatable-read snapshot, then map each block's coinbase payout address to its entity. Tx-graph coverage was complete: 6,582,885 non-coinbase inputs walked, 0 unresolved (each unresolved input would be a co-spend edge the clustering could not see).

The 50 to 12 collapse is driven entirely by that one cluster. Counterfactual (non-frozen; re-derived from the live index during the finding review): fusing only the 107-address cluster and no other merge already gives Nakamoto 12; applying every other merge but not that one gives 46.

2. Frozen versus non-frozen, and how each number was checked

Frozen (DS-004-entity-clustering-v1; sha256 in the MANIFEST):

• Every figure in the §1 table comes from summary.json and clusters.json, re-read from the frozen files on the recording date.
• Entity-basis Nakamoto 12 recomputes from clusters.json alone: sort the persisted entities by blocks and cumulate; the top 12 sum to 49,408 blocks, which exceeds 97,734 / 2 = 48,867, and the top 11 (48,471) do not. Rechecked against the frozen file on the recording date.
• Top-1 11.08% = 10,829 / 97,734; fleet share 9.14% = 8,931 / 97,734. Both denominators are the 97,734 blocks with a decoded coinbase payout address; recomputed from the frozen files on the recording date.

Non-frozen (API/Blockbook re-derivation; not reproducible from the frozen artifact):

• All per-address fleet figures in §3 (first-mined heights, per-address block counts, the ~29% share to height 30,768, the 10 lifetime spends, the 28.08M PRL received and its 98.2% residual). They were independently reproduced by the finding review against the live index and API, and are re-derivable there, but DS-004 does not contain them.
• The address basis (Nakamoto 50, effective 37.06) is frozen as a value in summary.json, but its re-derivation from the address distribution cannot be recomputed from the frozen artifact alone, because clusters.json persists only the top-100 entities (§7); it needs a live index.
• The §1 counterfactual (fleet-only merge already gives Nakamoto 12; every other merge but not the fleet gives 46) is NON-FROZEN: it is recorded only here, from the finding review's live-index re-derivation, not in any frozen artifact.

3. The 107-address fleet

Verified behavior (reproduced independently by the finding review; non-frozen per §2):

• All 107 addresses first mined within the chain's first 157 blocks.
• Per-address lifetime block counts are 66-108 (median 83): the signature of roughly 107 concurrently running, equally weighted workers started together.
• The fleet won roughly 29% of blocks up to height 30,768 (unit blocks; denominator canonical blocks up to that height; approximate, API-derived), then went dark: 0 blocks after 30,768 through the snapshot tip.
• The entire 107-address fusion rests on 10 lifetime non-coinbase spends: round integer-PRL principals to fresh one-use addresses, change to one shared address each time.
• 98.2% of the 28.08M PRL it received in coinbase payouts is still unspent (unit PRL; denominator the fleet's total coinbase receipts), about 27.6M PRL.

Inferred reading: this is single-wallet behavior. The alternative that would make the cluster an artifact rather than an entity, a shared service such as an exchange deposit path, predicts heterogeneous start times and volumes, frequent hot-wallet sweeps, and near-zero residual balances; every one of those signatures is absent here, and no labeled-pool address is fused anywhere in the run (multiPoolEntities = 0, frozen). The inference stops at one wallet controlling the payouts; §5 binds what may and may not be said beyond that.

Era qualifier (load-bearing): the fleet is a historical fact about the chain's first weeks. It stopped producing at height 30,768 and current-window concentration is unaffected. Any restatement of this finding must carry that qualifier; the all-time cumulative basis is the only basis on which the fleet ranks second.

4. What this bears on, and what it does not

• Revises: full-network all-time address-basis decentralization. The address count overstates the entity count, and the entity-basis Nakamoto of 12 is itself an upper bound on the true value, because clustering is a lower bound on linkage: merges the heuristic cannot see would only lower it further.
• Does not touch: the shipped /v1/pools labeled-pool concentration coefficient. Different unit (labeled pools, not clustered entities) and a 24h window, not all-time cumulative.
• Does not touch: OBS-005's 4.27 effective operators. That figure is over the 46 matched addresses of the weight-provenance scan on the DS-002-only basis (the combined basis is 48 addresses and 4.59; OBS-005 §10 item 5). Either way the matched addresses are disjoint from the 107 fleet addresses (checked by grep across the research trees on the run date).

5. Boundaries (binding)

1. Clusters are lower bounds on linkage, never identities. A coinbase paid straight into a shared custodial path would cluster onto the custodian; §3 records why that reading fails every signature here, but the ceiling on the claim is unchanged.
2. Report the fleet only as one wallet controlling 8,931 blocks' payouts. Never a named party, never a characterized party.
3. The heuristic cannot decide whether the key-holder owned the hashpower or coordinated it for others; an unlabeled private pool paying its workers from the same wallet looks identical on every measurement in this note. No claim either way is made.
4. The ~29% era share and all §3 per-address figures are approximate, API-derived, and non-frozen (§2); they are cited from the finding review's independent reproduction, not re-derived here.
5. Exploratory. No pre-registration constrains this analysis; it shipped as a roadmap phase (14.1) with a stated method and a frozen output, and its numbers should be treated accordingly.

6. Reproduction

• Frozen concentration figures: read datasets/DS-004-entity-clustering-v1/summary.json (sha256 pinned in the MANIFEST alongside the snapshot tip and hash).
• Entity-basis Nakamoto: from clusters.json, cumulate blocks over the persisted entities in descending order until the sum exceeds 48,867 (= 97,734 / 2); the crossing is at entity 12 (cumulative 49,408).
• Full re-run: keshictl cospend-cluster --out <new version dir> against a live index (cmd/keshictl/cospend.go; the tool refuses to overwrite an existing dataset directory, per the append-only rule).
• Fleet decomposition: non-frozen; re-derive per-address histories via the local Blockbook/API over the cluster's member addresses. Note that the membership list itself is not in the frozen artifact (only the representative address and counts are), so this path also needs the live index.

7. Data

datasets/DS-004-entity-clustering-v1/ (summary.json, clusters.json, MANIFEST.md), frozen, append-only. Known limitation, to fix in a future append-only version: clusters.json persists only the top-100 entities, so full cluster membership and the address-basis figures are not reproducible from the frozen artifact alone (both re-derive from a live index). A DS-004 v2 should persist full membership; corrections and extensions ship as a new version directory, never as an edit to v1.

8. Correction 2026-08-10: DS-004-v2 supersedes the v1 basis

Dated correction block per ERRATA-POLICY.md: an incomplete dataset is never edited in place; the fix is a new version directory plus this note. Everything above this section is the original v1-basis record and stays as written.

The new basis. The §7 limitation is closed by the append-only follow-up datasets/DS-004-entity-clustering-v2/ (generated 2026-08-10T07:20:27Z at snapshot tip 97,900, hash 526e0eb90df95ebfabccf9a6069857ad1799c6a9f132a4bbfe6c787c3c598407, REPEATABLE READ, tool 0705b6376718-we92ca47), which adds members.jsonl.gz: one JSON line per miner address ({address, entity, blocks}, sha256 pinned in the v2 MANIFEST). v2 is this note's basis from this date; v1 stays in place unedited.

Refreshed figures. Unit of analysis unchanged: blocks with a decoded coinbase payout address. The v2 snapshot holds 97,901 canonical blocks up to the bound, of which 97,900 carry one; those 97,900 blocks are the denominator of every share below, on both bases. Tx-graph coverage: 6,603,055 non-coinbase inputs walked, 0 unresolved. The v2 column is recomputed from the v2 committed files on the correction date (commands below); the v1 column is re-read from the v1 frozen files the same day.

What changed. The address-basis Nakamoto moved from 50 to 49, verified by recomputation from members.jsonl.gz: the top 49 addresses cumulate 48,982 blocks, which exceeds 97,900 / 2 = 48,950, and the top 48 (48,874) do not. §1's "falls from 50 (addresses) to 12 (entities)" therefore reads 49 to 12 on the v2 basis. The entity-basis Nakamoto stays 12 (top 12 cumulate 49,570; top 11 give 48,633). The fleet's block count is unchanged at 8,931; its share moves from 9.14% to 9.12% only because the denominator grew by 166 blocks. Every other shift is small drift from those 166 blocks and 2 additional addresses.

Previously non-frozen, now frozen-reproducible from the committed v2 artifact alone. Of §2's three non-frozen bullets, the second and third close fully and the first closes in part:

• The address-basis re-derivation (Nakamoto 49, effective units 36.90, top-1 11.09%). In v1 the values were stored in summary.json but could not be re-derived from the artifact; they now recompute from members.jsonl.gz.
• The §1 counterfactual: merging only the 107-address fleet and nothing else gives Nakamoto 12; applying every other merge but not the fleet gives 46. The same values the finding review derived from the live index on the v1 basis; both now reproduce from the committed file.
• The fleet's full membership (107 addresses, representative prl1p05tqqdf8tlzxlpt0vxmqp3ky9v76lh0egjejzjhcwwffq98j723q0gyl7t) and its per-address lifetime block counts: range 66-108, median 83, matching §3.

All of the above from the committed files, run at the repository root:

cd docs/research/datasets/DS-004-entity-clustering-v2 && python3 - <<'EOF'
import json, gzip, statistics
rows = [json.loads(l) for l in gzip.open('members.jsonl.gz', 'rt')]
tot = sum(r['blocks'] for r in rows)
ent = {}
for r in rows:
    ent.setdefault(r['entity'], []).append(r['blocks'])
def nakamoto(counts):
    s = 0
    for i, b in enumerate(sorted(counts, reverse=True), 1):
        s += b
        if s > tot / 2:
            return i, s
def eff(counts):
    return 1 / sum((b / tot) ** 2 for b in counts)
addr = [r['blocks'] for r in rows]
entc = [sum(v) for v in ent.values()]
print('addresses', len(addr), 'entities', len(ent), 'blocks', tot)
print('address basis: nakamoto %d (cum %d)' % nakamoto(addr),
      'eff %.2f' % eff(addr), 'top1 %d/%d = %.4f' % (max(addr), tot, max(addr) / tot))
print('entity basis:  nakamoto %d (cum %d)' % nakamoto(entc),
      'eff %.2f' % eff(entc), 'top1 %d/%d = %.4f' % (max(entc), tot, max(entc) / tot))
frep, fleet = next((k, v) for k, v in ent.items() if len(v) == 107)
print('fleet', frep, 'addresses', len(fleet), 'blocks', sum(fleet),
      'share %.4f' % (sum(fleet) / tot), 'per-address %d-%d median %g'
      % (min(fleet), max(fleet), statistics.median(fleet)))
only_fleet = [r['blocks'] for r in rows if r['entity'] != frep] + [sum(fleet)]
all_but = [sum(v) for k, v in ent.items() if k != frep] + fleet
print('counterfactuals: fleet-only merge nakamoto %d (cum %d);' % nakamoto(only_fleet),
      'all merges except the fleet nakamoto %d (cum %d)' % nakamoto(all_but))
EOF

Output on the correction date:

addresses 9561 entities 9303 blocks 97900
address basis: nakamoto 49 (cum 48982) eff 36.90 top1 10861/97900 = 0.1109
entity basis:  nakamoto 12 (cum 49570) eff 28.28 top1 10861/97900 = 0.1109
fleet prl1p05tqqdf8tlzxlpt0vxmqp3ky9v76lh0egjejzjhcwwffq98j723q0gyl7t addresses 107 blocks 8931 share 0.0912 per-address 66-108 median 83
counterfactuals: fleet-only merge nakamoto 12 (cum 49570); all merges except the fleet nakamoto 46 (cum 49044)

Still non-frozen, re-derived on the correction date against the local Blockbook (query-time tip 98,135), now walked over the frozen v2 membership instead of a live cluster listing. The walk ties itself to the artifact: the per-address coinbase counts equal the frozen blocks values for all 107 addresses, and the distinct coinbase txids total 8,931.

• Fleet first block at height 2, last at height 30,768; the latest per-address first-mined height is 157, matching §3's "first 157 blocks"; 0 fleet blocks above height 30,768 through the live tip.
• 10 lifetime non-coinbase spend txs (distinct txids with any fleet address on the input side), matching §3.
• Coinbase receipts 28,083,736 PRL; still unspent 27,580,034 PRL; unspent share 98.21% (unit PRL; denominator the fleet's total coinbase receipts), matching §3's ~98.2% of 28.08M PRL.
• The ~29% era share to height 30,768 keeps its §3 wording: approximate, live-derived, non-frozen.

Command (live index required):

cd docs/research/datasets/DS-004-entity-clustering-v2 && python3 - <<'EOF'
import json, gzip, urllib.request
def get(p):
    return json.load(urllib.request.urlopen('http://127.0.0.1:9131' + p, timeout=60))
rows = [json.loads(l) for l in gzip.open('members.jsonl.gz', 'rt')]
ent = {}
for r in rows:
    ent.setdefault(r['entity'], []).append(r)
fleet = next(v for v in ent.values() if len(v) == 107)
print('live tip at query time', get('/api/v2/')['blockbook']['bestHeight'])
heights, firsts, spends, cb_tx = [], [], set(), set()
recv = unspent = mismatches = 0
for m in fleet:
    a, txs, page, pages = m['address'], [], 1, 1
    while page <= pages:
        d = get('/api/v2/address/%s?details=txs&pageSize=1000&page=%d' % (a, page))
        pages = d.get('totalPages', 1); txs += d.get('transactions', []); page += 1
    mine = {}
    for t in txs:
        if any('coinbase' in v for v in t['vin']):
            val = sum(int(v['value']) for v in t['vout'] if a in (v.get('addresses') or []))
            if val: mine[t['txid']] = (t['blockHeight'], val)
        elif any(a in (v.get('addresses') or []) for v in t['vin']):
            spends.add(t['txid'])
    mismatches += (len(mine) != m['blocks'])
    hs = [h for h, _ in mine.values()]
    heights += hs; firsts.append(min(hs)); cb_tx |= set(mine)
    recv += sum(v for _, v in mine.values())
    unspent += sum(int(u['value']) for u in get('/api/v2/utxo/' + a) if u['txid'] in mine)
print('coinbase-count mismatches vs frozen members', mismatches,
      '; distinct coinbase txids', len(cb_tx))
print('first height', min(heights), '; last height', max(heights),
      '; max per-address first-mined', max(firsts),
      '; blocks above 30768:', sum(h > 30768 for h in heights))
print('lifetime non-coinbase spend txs', len(spends))
print('coinbase received %.2f PRL; unspent %.2f PRL; unspent share %.4f'
      % (recv / 1e8, unspent / 1e8, unspent / recv))
EOF

Output on the correction date:

live tip at query time 98135
coinbase-count mismatches vs frozen members 0 ; distinct coinbase txids 8931
first height 2 ; last height 30768 ; max per-address first-mined 157 ; blocks above 30768: 0
lifetime non-coinbase spend txs 10
coinbase received 28083736.35 PRL; unspent 27580034.05 PRL; unspent share 0.9821

§5's boundaries are unchanged and bind this correction as they bind the original text.